Skyvern demo: label simulated prompt + illustrative receipts (Tier 1)

[aimable100]

Summary

The Skyvern demo's attack mode scripts the agent (the prompt itself instructs it to pick ClearTone Ultra), and defended mode runs Tenuo authorization post-hoc against Skyvern's output rather than inline at each browser action. Both were plausible to a developer reading the code, but the blog/tutorial framed them as real inline interception. This PR closes the gap without changing the architecture — Tier 1 of the credibility-strategy doc. The defense story (compromised LLM, correct outcome, signed receipts) is unchanged; the framing now matches the code.

• Runner — rename TASK_PROMPT_ATTACK → TASK_PROMPT_ATTACK_SIMULATED. Add --live flag that runs attack/defended with the neutral comparison prompt so the DOM injection alone has to persuade the model. Tag the redirect/fallback/checkout receipts with synthetic=True; print_receipt_chain renders them as [ILLUSTRATIVE] with a header explaining the post-hoc reconstruction. Mode-aware "agent tricked" messaging.
• Blog — new "How to read this demo" callout up front (simulated vs live, post-hoc disclosure), mode-aware Act 2 framing, [ILLUSTRATIVE] tags on receipts 7/8/9 with reasons, and a note under the architecture diagram about post-hoc vs Tier-2 inline. Fixes the browser_extract OneOf([...]) vs Wildcard() drift between blog and code.
• Tutorial — documents --simulated vs --live; calls out post-hoc; labels illustrative receipts; "LLM tricked?" → "LLM compromised?" in the comparison table.
• demo-outputs — EDITED TRANSCRIPT — NOT A RAW RUN LOG header on all three files so the styled blog excerpts aren't mistaken for raw captures.

Tier 2 (inline integration at Skyvern's action handler) is intentionally not in this PR — that's a separate effort that requires forking/monkeypatching Skyvern.

Test plan

• python skyvern-config/task.py clean — happy path unchanged
• python skyvern-config/task.py attack — prints [SIMULATED] banner; selects ClearTone Ultra; ends with [SIMULATED] Agent selected ClearTone Ultra as scripted by the prompt
• python skyvern-config/task.py attack --live — prints [LIVE] banner; neutral prompt; outcome depends on model
• python skyvern-config/task.py defended --local — receipt chain shows [ILLUSTRATIVE] tags on receipts 7/8/9 and the post-hoc header
• python skyvern-config/task.py --help reads cleanly and explains the modes
• Blog and tutorial render correctly on GitHub (callouts, code blocks, table)